Research On Application Of Bayesian Net In Robust Speech Recognition

Automatic Speech Recognition (ASR for short) is a high-technique which can transform speech signal to corresponding texts or commands. In the past few years, ASR has achieved a great success in laboratory. However, in practical application, the recognition environment is quite different from the training one, which is called mismatch. Because of the mismatch, the recognition system deteriorates seriously. In order to make the recognition system practical, researchers have to try their best to minish the impact which the mismatch makes on recognition system.A common technique for robust speech recognition is feature compensation. This thesis compensates speech features based on Bayesian net theory, which is flexible in modeling and has a simple but effective learning algorithm—VBEM.The features that are compensated in this thesis are energy and Mel-Frequency cepstrum coefficients. Two methods are used to compensate energy feature. The first chooses RASTA-PLP energy which is estimated using MMSE instead of spectrum energy as energy feature. In 10dB SNR white noise environment, when compared to systems with no energy compensation modules, this method improves speech recognition system accuracy by 2.82%. The second compensates the spectrum energy with the learning algorithm of Bayesian net, and makes a excellent estimation of spectrum energy. The speech recognition system accuracy is improved by 4.21% in 10dB SNR white noise environment.The method for compensating Mel-Frequency cepstrum coefficients is based on Algonquin framework. This method fuses energy feature using Bayesian net theory, then, in 10dB SNR white noise environment, the speech recognition system accuracy is improved by 2.24% whencompared with Algonquin.